Yukti
v0.5.1air-gappedSIH 2026 · SIH26117

Plant knowledge,
kept inside.

cited · authorized
Windows 10 / 11macOS 13.3+ · Apple siliconworks offline source
01
~0ms
Laya routes before any LLM call
02
0–0tok/s
Gemma 3 4B · 6 GB laptop GPU
03
Enहिंಕ
English · Hindi · Kannada, offline speech
04
0bytes out
Offline guard blocks and counts egress
0102:15 AM

One trip. Six systems.

PUMP A202:15
Equipment master
Maintenance history
Drawings · P&IDs
SOP rev 2 vs rev 3
Scanned reports
Emails · shift logs
A2 dossierallowed
  • Equipment master[S1]
  • Maintenance history[S2]
  • Drawings · P&IDs[S3]
  • SOP rev 2 vs rev 3[S4]conflicting
  • Scanned reports[S5]
  • Emails · shift logs[S6]
02End to end

From a question to a decision.

    1
    Question or scan
    text · photo · voice
    2
    Identity & policy
    post · dept · clearance · scope
    3
    Laya routes
    typed decision · ~1 ms
    4
    Durable queue
    specialist worker
    5
    Local LLM
    OCR + VLM · analytics · sandbox
    6
    Cited evidence
    known · missing · conflicting
    7
    Answer · note · alert
    accountable human
03Trust

An answer the right person is allowed to trust.

S1Evidence
sourceSOP-EL-014
revisionrev 3CURRENTrev 2
page · recordp.2·A2_WorkOrders!34
timestamp2026-09-29
confidence0.84
status● KNOWN○ MISSING◐ CONFLICTING
found bybm25tagsgraphvector
Policy engine
ABAC · deny-overrides · default deny
  1. ElectricianA2 dossierallowed
  2. Boiler engineerFinance auditdenied
  3. Boiler engineerAccess request · whyto manager
  4. Dept. managerTime-bound grant≤ 72 h
every decision → SHA-256 audit chain
01
The LLM never grants access
host policy · before retrieval & every tool
02
Small model decides first
denials & off-topic answered without the LLM
03
Advisory only
no tool writes to DCS / SCADA
04Features

One workbench. Twelve instruments.

Laya · ONNX
  • policy
  • small LLM
  • OCR + VLM
  • analytics
  • 27B reasoning
  • coder
  • workflow
Laya + work queuepicks the worker
bm25
asset tags
metadata
knowledge graph
BGE-M3 vectors
behind the policy check
Hybrid retrieval
.docx
.pptx
.xlsx
steps
$ sandbox.run(calc.py) ✓ · no network
Agentic workbench
OCR + VLMverifyreviewfindingnotelane
Scan → finding
electrical
mechanical
HSE
procurement
finance
management
acknowledgeescalatedecide
Approver lanesacknowledge · escalate · follow up
P-2041A certok
PSV-118 intervalok
M-A2 warrantyexpires soon
Asset ledgeralerts before lapse
forecast ± bandHiGHS LPplanners decide
Production intelligence
5.1of 6 GB
CUDA / VulkanVRAM auto-fitq8_0 KV cachehot KV slotsspeculative
Performance Engine
TAG
P-2041A
matchedrust
Ask with a photoadd-on
Enहिंदीಕನ್ನಡ
Speak instead of typingadd-on
isolated
no docs · no tools · no approvals
Isolated research
140public sources
Knowledge basecontacts · history · records
EXAMPLE
plant documents · written by Team UniMinds
PUBLIC · cited
MRPL public facts · source links
Real vs example datawatermarked in the UI
05Architecture

Four layers. One server. Your plant.

Desktop app
Browser
GPU server
    1
    Who is asking
    argon2id + MFA · sessions · IAM attributes
    2
    Decide & route
    ABAC policy · Laya · ONNX · durable queue
    3
    Knowledge & output
    hybrid retrieval · facts · ledger · workbench
    4
    Performance Engine
    llama.cpp · CUDA / Vulkan · hot KV
#0001policy.deny10800001…
#0002retrieval20080808…
#0003tool.call30000000…
#0004chat.answer40800000…
#0005approval50000800…
SHA-256 · hash-chained · verifiable
PythonFastAPIReact + ViteElectronllama.cppONNX RuntimeCUDAVulkanGemma 3 4BRapidOCRwhisper.cppAI4BharatHiGHS LPSQLite FTS5BGE-M3argon2idABACSHA-256
PythonFastAPIReact + ViteElectronllama.cppONNX RuntimeCUDAVulkanGemma 3 4BRapidOCRwhisper.cppAI4BharatHiGHS LPSQLite FTS5BGE-M3argon2idABACSHA-256
06Why on-prem

Cloud AI is off-limits for P&IDs.

Cloud assistants
Yukti
Data
leaves the plant
on-premise · air-gapped
Cost
pay per token
one-time hardware
Serving
generic
q8_0 KV · speculative · profiles
Trust
fluent, unsourced
cited · known · missing · conflicting
07Measured

No cluster. No cloud. A 6 GB laptop GPU.

tok/sGemma 2 2B Q8_0RTX 4050 · 6 GB
llama.cpp CUDA 12.43.9× CPU51.7
llama.cpp Vulkan50.8
CPU only13.4
Gemma 3 4B · 44–58 tok/s
0.2 s
first token
~1 ms
Laya routing
4–10 s
photo question
0.3 s
speech → text
5.1 / 6 GB
Gemma + Whisper VRAM
08Roles

Ten roles. Need-to-know, by design.

  • Employee
  • Head of Dept.
  • Approver lanes
  • Plant manager
  • Top management
  • Planner
  • Stores
  • Auditor
  • Researcher
  • Administrator
09Security

Five walls around every answer.

LLM never decides access
deny-overrides · default deny · fail closed
Offline guard
blocks & counts outbound connections
Tamper-evident audit
logins · denials · tools · answers · approvals
Advisory only
no writes to DCS / SCADA
Sandbox & isolation
code without network · research walled off
10Get started

One download. Windows or Mac.

01
One file
Yukti-Setup.exe · Yukti-macOS-arm64.dmg
02
Open it
Windows: double-click · Mac: drag to Applications
03
Downloads its AI once
3.9 GB · 5.3 GB NVIDIA · photos + voice included · verified
04
Sign in
works offline
Photos~800 MB · included
Voice0.3–1 GB · included
After that0 bytes online
11Roadmap

Pilot first. Then the plant.

phase 1
Pilot
one unit · dossiers · SOPs · drawings · cited RAG
phase 2
Workflows
scan → finding · approvals · ledger · alerts
phase 3
Planning
production intelligence with planners
phase 4
Scale
plant-wide · template for refineries & PSUs
nextcode-signed installersplant PKI certificatesrefinery guard modelDMS · CMMS · SAP connectorsLinux serverevaluation on real documents

Bring Yukti inside.

v0.5.1·Windows 10 / 11·macOS 13.3+ · Apple silicon·all releases·docs
Windows: SmartScreen → More info → Run anyway | Mac: Privacy & Security → Open Anyway | not yet code-signed